[jeckersb]

The Koji RPM build was spending a lot of time recompiling the same code
over and over. This series removes those duplicate builds:

• Drop panic = "abort" from the release profile. Since Rust 1.81
  (we need at least 1.85), a panic that reaches an extern "C" boundary
  aborts on its own, so this setting is no longer needed for safety. The
  test harness can't use it, so cargo test --release / %cargo_test
  was rebuilding every dependency with a different panic setting.
• Enable tokio test-util in the workspace dependency. Before, only
  a dev-dependency turned it on, so tokio and everything that depends on
  it was compiled once for build and again for test. With this change,
  cargo test --release after cargo build --release rebuilds 20 crates
  instead of 36. The binary grows by ~28KB (0.1%).
• Don't rebuild during RPM %install. make install re-ran cargo
  three times (~12 minutes in Koji). These --release builds also wrote
  over the %cargo_build output through the target/release ->
  target/rpm symlink, so the RPM shipped binaries built with our
  release profile instead of Fedora's rpm profile. Add SKIP_BUILD=1
  to the Makefile and use it from the spec. Plain make install works
  the same as before.
• Don't build bootc to generate man pages in the RPM. CI already
  checks that the generated sections of the man pages match the CLI, so
  re-syncing them in a release build changes nothing and costs a full
  extra build (~7 minutes). Add xtask manpages --no-sync /
  make manpages-no-sync and build xtask with the rpm profile and
  %{__cargo} so it doesn't trigger a rebuild in %cargo_test.
  The output is identical to make manpages.

Assisted-by: AI

[Copilot]

🔵 Needs a closer look

The non-container build’s new rpm Cargo profile assumption needs maintainer validation for older supported RHEL 9 rust-toolsets because the macro definitions needed to check legacy-branch compatibility are absent from this snapshot.

No open issues 🎉

What changed in this PR

This PR reduces redundant Cargo compilation during RPM packaging while preserving the binaries built with Fedora’s rpm profile and keeping ordinary make install behavior unchanged.

Changes:

• Align release and test builds by removing panic = "abort" and enabling Tokio’s test-util feature at workspace level.
• Add SKIP_BUILD=1 so RPM installation reuses existing build artifacts.
• Generate man pages without rebuilding bootc, using the RPM Cargo environment and profile for xtask.

File	Description
Makefile	Adds build-skipping controls, configurable Cargo invocation, and a no-sync man-page target.
crates/​xtask/​src/​xtask.rs	Adds and dispatches the manpages --no-sync option.
crates/​xtask/​src/​man.rs	Adds man-page generation from existing Markdown without CLI synchronization.
crates/​lib/​Cargo.toml	Removes the redundant Tokio dev-dependency configuration.
contrib/​packaging/​bootc.spec	Uses no-sync man-page generation and skips rebuilding during installation.
Cargo.toml	Removes release panic-abort behavior and enables Tokio test-util workspace-wide.

_{Review effort: Balanced}

[jeckersb]

🔵 Needs a closer look

The non-container build’s new rpm Cargo profile assumption needs maintainer validation for older supported RHEL 9 rust-toolsets because the macro definitions needed to check legacy-branch compatibility are absent from this snapshot.

Good to consider that. Poked around a bit with an agent and came up with this:

Checked macros.rust-toolset from el9 rust-toolset 1.84.1, 1.92.0 and
1.99.1: all define [profile.rpm], create the target/release →
target/rpm symlink in %cargo_prep, and use --profile rpm in
%cargo_build/%cargo_test. 1.75.0 (9.4) predates this, but it is
already unsupported by main (MSRV 1.85, and the spec needs %cargo_prep
-N and %cargo_vendor_manifest).

So I think this is fine as-is

[jeckersb]

ballpark comparison, over here the f44 package job took 31m20s whereas the same job here took 19m47s. But CI uses the "container" packaging flow which previously was already a bit better optimized than the "real" builds in koji. For me a local mockbuild went from about 15m down to 5m.

[jmarrero]

@jmarrero-bot review this and explain how is the rpm produced with this change different from the current ones.

[jmarrero-bot]

@jmarrero The packaging changes look good to merge. Two things are worth a decision first: dropping panic = "abort", and where tokio test-util is enabled. Reviewed at a7f0421.

I built both RPMs Koji-style (%cargo_build, %check on) in fedora:44 on 16 cores, base 709dfe5 vs head:

• rpmbuild takes 12m18s instead of 26m49s: %install drops from 9m12s to 21s, and %check rebuilds 20 crates instead of 358.
• Confirmed: today's RPM ships the release-profile bootc that %install relinks through the target/release -> target/rpm symlink. On Fedora the codegen flags were effectively the same (%install gets the same RUSTFLAGS, and profile.rpm inherits release). What was missing is RUSTC_BOOTSTRAP=1, so anyhow/thiserror/cap-std were built without their nightly cfgs (error_generic_member_access, can_vector, write_all_vectored, io_error_more). Those only affect backtrace forwarding and vectored-I/O hints, so bootc behaves the same either way. On EL9, %install exports no RUSTFLAGS at all, so the old binary there used codegen-units=16.
• Apart from the version string in LICENSE.dependencies, only the three binaries (and their debuginfo) differ. Man pages, units, completions, configs and Requires/Provides are byte-identical.
• /usr/bin/bootc grows from 19.08 to 22.81 MB (+19.5%; the bootc RPM goes from 5.9 to 7.3 MB): about 2.6 MB of landing-pad code in .text plus 1.1 MB of unwind tables. A forced panic (bootc completion bash > /dev/full) now exits 101 with no core dump, instead of SIGABRT with a core.

Findings:

• Cargo.toml:8: dropping panic = "abort" accounts for nearly all of the ~3.7 MB per image; together with test-util it saves about 1 minute of %check here (16 cores). The "~28KB" in the description covers only test-util. The FFI reasoning holds, so this is a size and crash-reporting trade-off.
• Cargo.toml:77: with test-util on the workspace tokio dependency, cargo package writes it into the published bootc-internal-utils and bootc-internal-blockdev manifests, so bootupd would get it too. Enabling it on an unpublished crate's tokio dependency (e.g. crates/cli) should avoid the rebuild just as well.

Commands, logs, readelf output and tables: https://gist.github.com/jmarrero-bot/23ecd97acbd9cabb70d4f0c56f8092c9

Generated-by: https://github.com/jmarrero-bot#llms